An article information pushing method and device, electronic equipment and storage medium

Abstract

Embodiments of the present application provide an article information pushing method and device, electronic equipment and storage medium, by obtaining the first user feature vector for the recommended user and the recommended article feature vector for the recommended article; determining the source model, and based on the source model, calculating the basic matching probability for the recommended user and the recommended article through the first user feature vector and the recommended article feature vector; obtaining the second user feature vector for the recommender user; training the source model through the first user feature vector, the second user feature vector and the basic matching probability to generate the target model; and pushing the article information for the recommended article by using the target model, so that the target model can push the article information to the recommended user based on the relevance of the recommender user and the recommended user when pushing the article information, and further improve the efficiency of pushing the article information to the user through the deep learning model.

Description

Technical Field

[0001] This invention relates to the field of item information push technology, and in particular to an item information push method, an item information push device, an electronic device, and a computer-readable storage medium. Background Technology

[0002] Telemarketing refers to using telephone operators to attract new customers and contact existing customers to determine their satisfaction or willingness to place an order. With the development of technology, telemarketing using deep learning models has gradually replaced manual telephone sales. Using deep learning models for telemarketing can improve efficiency and reduce costs.

[0003] However, when using deep learning models for telemarketing, the technology only adaptively trains the model based on customer and product characteristics. This results in a low degree of compatibility between the trained model and the customer, thus reducing the efficiency of pushing product information. Summary of the Invention

[0004] The present invention provides a method, apparatus, electronic device, and computer-readable storage medium for pushing item information to users, in order to solve the problem of how to improve the efficiency of pushing item information to users through deep learning models.

[0005] This invention discloses a method for pushing item information, including:

[0006] Obtain the first user feature vector for the recommended user and the recommended item feature vector for the recommended item;

[0007] A source model is determined, and based on the source model, a basic matching probability for the recommended user and the recommended item is calculated using the first user feature vector and the recommended item feature vector.

[0008] Obtain the second user feature vector for the recommender's users;

[0009] The source model is trained using the first user feature vector, the second user feature vector, and the basic matching probability to generate the target model;

[0010] The target model is used to push item information for the recommended items.

[0011] Optionally, the step of obtaining the first user feature vector for the recommended user and the recommended item feature vector for the recommended item may include:

[0012] Obtain structured and unstructured user feature vectors for the recommended users;

[0013] Obtain the structured feature vector and unstructured feature vector of the recommended items.

[0014] Optionally, the step of calculating the basic matching probability for the recommended user and the recommended item based on the source model using the first user feature vector and the recommended item feature vector may include:

[0015] The structured user feature vector is input into the source model to calculate and generate the first feature loss rate;

[0016] The structured recommended item feature vector is input into the source model to calculate and generate the second feature loss rate;

[0017] The unstructured user feature vector is input into the source model to calculate and generate the third feature loss rate;

[0018] The unstructured recommended item feature vector is input into the source model to calculate and generate the fourth feature loss rate;

[0019] When the first feature loss rate, the second feature loss rate, the third feature loss rate, and the fourth feature loss rate are all below a preset threshold, the structured user feature vector, the unstructured user feature vector, the structured recommended item feature vector, and the unstructured item feature vector are used to calculate and generate the basic matching probability for the recommended user and the recommended item through the source model.

[0020] Optionally, the source model has a corresponding source domain, and the step of training the source model using the first user feature vector, the second user feature vector, and the basic matching probability to generate the target model may include:

[0021] The source model is adaptively trained using the basic matching probability and the second user feature vector to generate an initial target model; the initial target model has a corresponding target domain.

[0022] The first user feature vector and the second user feature vector are shared to the target domain, and when it is determined that there is a first mapping relationship between the source domain feature vector in the source domain and the target domain feature vector in the target domain, the target domain error value for the target domain is obtained.

[0023] When the target domain error value is less than a preset threshold, the initial target model is determined as the target model.

[0024] Optionally, it may also include:

[0025] When the target domain error value is greater than or equal to a preset threshold, a second mapping relationship is constructed between the source domain feature vector in the source domain and the target domain feature vector in the target domain; the second mapping relationship is used to adjust the target domain error value.

[0026] This invention also discloses an item information push device, comprising:

[0027] The feature vector acquisition module is used to acquire the first user feature vector for the recommended user and the recommended item feature vector for the recommended item.

[0028] The basic matching probability calculation module is used to determine the source model and, based on the source model, calculate and generate the basic matching probability for the recommended user and the recommended item using the first user feature vector and the recommended item feature vector.

[0029] The second user feature vector acquisition module is used to acquire the second user feature vector for the recommender user.

[0030] The target model generation module is used to train the source model using the first user feature vector, the second user feature vector, and the basic matching probability to generate a target model.

[0031] The item information recommendation module is used to push item information for the recommended items using the target model.

[0032] Optionally, the feature vector acquisition module includes:

[0033] The user feature vector acquisition submodule is used to acquire structured and unstructured user feature vectors for the recommended users.

[0034] The item feature vector acquisition submodule is used to obtain structured and unstructured item feature vectors for recommended items.

[0035] Optionally, the basic matching probability calculation module may include:

[0036] The first feature loss rate calculation submodule is used to input the structured user feature vector into the source model and calculate and generate the first feature loss rate;

[0037] The second feature loss rate calculation submodule is used to input the structured recommended item feature vector into the source model and calculate and generate the second feature loss rate.

[0038] The third feature loss rate calculation submodule is used to input the unstructured user feature vector into the source model and calculate the third feature loss rate.

[0039] The fourth feature loss rate calculation submodule is used to input the unstructured recommended item feature vector into the source model and calculate and generate the fourth feature loss rate.

[0040] The basic matching probability calculation submodule is used to calculate and generate a basic matching probability for the recommended user and the recommended item by using the structured user feature vector, the unstructured user feature vector, the structured recommended item feature vector, and the unstructured item feature vector through the source model when the first feature loss rate, the second feature loss rate, the third feature loss rate, and the fourth feature loss rate are lower than a preset threshold.

[0041] Optionally, the source model has a corresponding source domain, and the target model generation module may include:

[0042] An initial target model generation submodule is used to perform feature adaptive training on the source model using the basic matching probability and the second user feature vector to generate an initial target model; the initial target model has a corresponding target domain.

[0043] The target domain error value acquisition submodule is used to share the first user feature vector and the second user feature vector to the target domain, and when it is determined that there is a first mapping relationship between the source domain feature vector in the source domain and the target domain feature vector in the target domain, the target domain error value for the target domain is acquired.

[0044] The target model determination submodule is used to determine the initial target model as the target model when the target domain error value is less than a preset threshold.

[0045] Optionally, it may also include:

[0046] The second mapping relationship construction submodule is used to construct a second mapping relationship between the source domain feature vector in the source domain and the target domain feature vector in the target domain when the target domain error value is greater than or equal to a preset threshold; the second mapping relationship is used to adjust the target domain error value.

[0047] This invention also discloses an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0048] The memory is used to store computer programs;

[0049] When the processor executes a program stored in the memory, it implements the method described in the embodiments of the present invention.

[0050] This invention also discloses a computer-readable storage medium storing instructions that, when executed by one or more processors, cause the processors to perform the methods described in this invention.

[0051] The embodiments of the present invention have the following advantages:

[0052] In this embodiment of the invention, a first user feature vector for the recommended user and a recommended item feature vector for the recommended item are obtained; a source model is determined, and based on the source model, a basic matching probability for the recommended user and the recommended item is calculated using the first user feature vector and the recommended item feature vector; a second user feature vector for the recommending user is obtained; the source model is trained using the first user feature vector, the second user feature vector, and the basic matching probability to generate a target model; and the target model is used to push item information for the recommended item. This allows the target model to push item information to the recommended user based on the correlation between the recommending user and the recommended user, thereby improving the efficiency of pushing item information to users through a deep learning model. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the steps of an item information push method provided in an embodiment of the present invention;

[0054] Figure 2 This is a flowchart illustrating an item information push method provided in an embodiment of the present invention;

[0055] Figure 3 This is a structural block diagram of an item information push device provided in an embodiment of the present invention;

[0056] Figure 4 This is a hardware structure block diagram of an electronic device provided in various embodiments of the present invention. Detailed Implementation

[0057] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0058] In practical applications, existing recommendation models mostly encode customer characteristics (such as needs, interests, preferences, consumption behavior, browsing history, etc.) and product characteristics, generating recommendation lists or bundled recommendations through methods such as combined mining or matrix model scoring mechanisms to achieve a one-to-one optimal match between customers and products. The model's adaptive adjustment also involves analyzing and optimizing data features such as customer behavior patterns and characteristics within the internet system. However, if the trained model is considered as a "recommender," it fails to consider the "recommender" factor. For scenarios relying on manual methods (such as customer service representatives) for product recommendations, such as telemarketing, it cannot adaptively adjust based on the differences in the human recommendation interaction process. Furthermore, these technologies do not consider the multi-task relevance of human recommendations; that is, they do not consider factors such as user needs, company interests, and recommender benefits, thus failing to fundamentally recommend products that suit user needs and satisfy the needs of all parties.

[0059] Reference Figure 1 The diagram illustrates a flowchart of a method for pushing item information according to an embodiment of the present invention, which may specifically include the following steps:

[0060] Step 101: Obtain the first user feature vector for the recommended user and the recommended item feature vector for the recommended item;

[0061] Step 102: Determine the source model, and based on the source model, calculate and generate the basic matching probability for the recommended user and the recommended item using the first user feature vector and the recommended item feature vector;

[0062] Step 103: Obtain the second user feature vector for the recommender user;

[0063] Step 104: Train the source model using the first user feature vector, the second user feature vector, and the basic matching probability to generate the target model;

[0064] Step 105: Use the target model to push item information for the recommended items.

[0065] In a specific implementation, embodiments of the present invention can obtain a first user feature vector for the recommended user and a recommended item feature vector for the recommended item. Specifically, the recommended user can be a user who receives item information, such as a customer who is recommended a product in a telephone sales call, and the product recommended in a telephone sales call can be a recommended item.

[0066] In practical applications, the user being recommended can possess corresponding user characteristic information, such as age, gender, needs, interests, preferences, consumption behavior, and browsing history. The second user feature vector can then be a vector used to express the user characteristic information of the recommending user. Similarly, recommended items can possess corresponding recommended item characteristic information. For example, in telemarketing, the type of virtual package and the services included in the package, such as call duration, number of SMS messages, and data usage, can be considered. The recommended item feature vector can then be a vector used to express the recommended item's characteristic information.

[0067] The embodiments of the present invention can determine the source model. Specifically, the source model may include, but is not limited to, multi-task models such as ESMM model and MMOE model.

[0068] After determining the source model, embodiments of the present invention can calculate and generate the basic matching probability of the recommended user and the recommended item based on the source model using the first user feature vector and the recommended item feature vector.

[0069] For example, a vector representing the user features of the recommending user and a vector representing the features of the recommended item can be input into an ESMM model or an MMOE model to enable adaptive training of the ESMM model or MMOE model, thereby outputting a basic matching probability for the recommending user and the recommended item. This basic matching probability can be used to express the degree of fit between the recommending user and the recommended item. That is, the higher the basic matching probability, the higher the degree of fit between the recommending user and the recommended item.

[0070] The embodiments of the present invention can obtain a second user feature vector for the recommending user. In practical applications, the recommending user can be the user who pushes product information, such as the salesperson who recommends products in telephone sales. The recommending user can have corresponding second user feature information, such as performance information, personality feature information, historical performance information, service skill evaluation information, etc. The second user feature vector can be a vector used to express the second user feature information.

[0071] After obtaining the first user feature vector, the second user feature vector, and the recommended item feature vector, and calculating the basic matching probability between the recommended user and the recommended item, this embodiment of the invention can train a source model using the first user feature vector, the second user feature vector, and the basic matching probability to generate a target model. The target model can then be used to push item information for the recommended item, thereby enabling the target model to push item information to the recommended user based on the correlation between the recommending user and the recommended user when pushing item information.

[0072] In this embodiment of the invention, a first user feature vector for the recommended user and a recommended item feature vector for the recommended item are obtained; a source model is determined, and based on the source model, a basic matching probability for the recommended user and the recommended item is calculated using the first user feature vector and the recommended item feature vector; a second user feature vector for the recommending user is obtained; the source model is trained using the first user feature vector, the second user feature vector, and the basic matching probability to generate a target model; and the target model is used to push item information for the recommended item. This allows the target model to push item information to the recommended user based on the correlation between the recommending user and the recommended user, thereby improving the efficiency of pushing item information to users through a deep learning model.

[0073] Meanwhile, because the target model pushes item information to the recommended user based on the correlation between the recommender and the recommended user, it can provide more targeted item information push services to the recommended user in this process, thereby improving the user experience of the recommended user when receiving item information.

[0074] Based on the above embodiments, modified embodiments of the above embodiments are proposed. It should be noted that, in order to keep the description brief, only the differences from the above embodiments are described in the modified embodiments.

[0075] In an optional embodiment of the present invention, the step of obtaining the first user feature vector for the recommended user and the recommended item feature vector for the recommended item includes:

[0076] Obtain structured and unstructured user feature vectors for the recommended users;

[0077] Obtain the structured feature vector and unstructured feature vector of the recommended items.

[0078] In practical implementation, with the massive generation of unstructured data, using only structured data for model training no longer meets current market demands. Structured and unstructured data are two types of big data. Structured data can be highly organized and neatly formatted data, such as data types placed in tables and spreadsheets. Structured data is also known as quantitative data, which is information that can be represented by data or a unified structure, such as numbers and symbols. Typical structured data can include, but is not limited to, dates, phone numbers, addresses, product names, etc. Unstructured data is essentially all data other than structured data. It does not conform to any predefined model, so it is stored in non-relational databases. Unstructured data can be text or non-text, and it can be human-generated or machine-generated. Unstructured data can also be data with variable fields, such as features like tone of voice, emotion, energy, and silence duration.

[0079] The structured user feature vectors in this embodiment of the invention can be used to express the structured data of the recommended user, the unstructured user feature vectors can be used to express the unstructured data of the recommended user, the structured recommended item feature vectors can be used to express the structured data of the recommended item, and the unstructured item feature vectors can be used to express the unstructured data of the recommended item.

[0080] The embodiments of the present invention can obtain structured user feature vectors and unstructured user feature vectors for the recommended user, and obtain structured recommended item feature vectors and unstructured item feature vectors for the recommended item.

[0081] In the specific implementation, to obtain structured user feature vectors and structured recommended item feature vectors, continuous variables can be preprocessed. For example, the units of measurement can be standardized to ensure that dimensions do not become invalid due to unit differences. Additionally, missing values ​​can be filled in appropriately for different dimensions. Then, discrete variables can be preprocessed, such as time-series dimensions and gender, using OneHot encoding. This allows for the acquisition of structured user feature vectors and structured recommended item feature vectors.

[0082] To obtain unstructured user feature vectors and unstructured recommended item feature vectors, a feature extraction model can be built for text data, mainly audio recordings and work order descriptions. The Word2Vec method is primarily used for extraction, and the BERT model is mainly used for analysis. Then, a deep learning feature extraction model can be built for audio data to extract and label features such as tone of voice, emotion, energy, and silence duration. Word2vec is a group of models used to generate word vectors. These models are shallow, two-layer neural networks used for training to reconstruct linguistic word text. The network represents words and needs to guess the input words in adjacent positions. Under the bag-of-words assumption in word2vec, the order of words is unimportant. After training, the word2vec model can be used to map each word to a vector, which can be used to represent the relationship between words. This vector is the hidden layer of the neural network. BERT, from a Google paper, is a pre-training of deep bidirectional transformers for language understanding. BERT is an acronym for "Bidirectional Encoder Representations from Transformers". Overall, BERT is an autoencoder language model.

[0083] In this embodiment of the invention, by obtaining structured user feature vectors and unstructured user feature vectors for the recommended user, and structured recommended item feature vectors and unstructured item feature vectors for the recommended items, the first user feature vector and the recommended item feature vector are extracted from the perspectives of structured data and unstructured data respectively, thereby further improving the accuracy and efficiency of pushing item information to users through deep learning models.

[0084] In an optional embodiment of the present invention, the step of calculating the basic matching probability for the recommended user and the recommended item based on the source model using the first user feature vector and the recommended item feature vector includes:

[0085] The structured user feature vector is input into the source model to calculate and generate the first feature loss rate;

[0086] The structured recommended item feature vector is input into the source model to calculate and generate the second feature loss rate;

[0087] The unstructured user feature vector is input into the source model to calculate and generate the third feature loss rate;

[0088] The unstructured recommended item feature vector is input into the source model to calculate and generate the fourth feature loss rate;

[0089] When the first feature loss rate, the second feature loss rate, the third feature loss rate, and the fourth feature loss rate are all below a preset threshold, the structured user feature vector, the unstructured user feature vector, the structured recommended item feature vector, and the unstructured item feature vector are used to calculate and generate the basic matching probability for the recommended user and the recommended item through the source model.

[0090] In this embodiment of the invention, a structured user feature vector can be input into the source model to calculate a first feature loss rate, a structured recommended item feature vector can be input into the source model to calculate a second feature loss rate, an unstructured user feature vector can be input into the source model to calculate a third feature loss rate, and an unstructured recommended item feature vector can be input into the source model to calculate a fourth feature loss rate.

[0091] Optionally, when the first feature loss rate and the second feature loss rate are higher than or equal to a preset threshold, the feature extraction method of the structured user feature vector or the structured recommended item feature vector can be adjusted, and the model parameters can be fine-tuned to make the feature loss rate lower than the preset threshold; when the third feature loss rate and the fourth feature loss rate are higher than or equal to the preset threshold, the feature extraction method of the unstructured user feature vector or the unstructured recommended item feature vector can be adjusted, and the feature extraction model parameters can be fine-tuned to make the feature loss rate lower than the preset threshold.

[0092] When the loss rates of the first, second, third, and fourth features are all below a preset threshold, the basic matching probabilities for the recommended user and the recommended item are calculated using the source model by employing structured user feature vectors, unstructured user feature vectors, structured recommended item feature vectors, and unstructured item feature vectors.

[0093] For example, an algorithm group with high accuracy and precision after feature learning can be selected as the core algorithm of the source model, such as the GBDT (Gradient Boosting Decision Tree) algorithm, the Random Forest (RF) algorithm, etc. Reasonable algorithm selection and integration rules are set, and a multi-task learning model is introduced as the source model. The model is trained to understand users based on company revenue and user needs. The source model is used to calculate and generate the basic matching probability for the recommended user and the recommended item.

[0094] In this embodiment of the invention, a first feature loss rate is calculated by inputting the structured user feature vector into the source model; a second feature loss rate is calculated by inputting the structured recommended item feature vector into the source model; a third feature loss rate is calculated by inputting the unstructured user feature vector into the source model; and a fourth feature loss rate is calculated by inputting the unstructured recommended item feature vector into the source model. When the first, second, third, and fourth feature loss rates are all lower than preset thresholds, the basic matching probability for the recommended user and the recommended item is calculated by the source model using the structured user feature vector, the unstructured user feature vector, the structured recommended item feature vector, and the unstructured item feature vector. This improves the accuracy of the basic matching probability and further enhances the accuracy and efficiency of pushing item information to users through the deep learning model.

[0095] In an optional embodiment of the present invention, the source model has a corresponding source domain, and the step of training the source model using the first user feature vector, the second user feature vector, and the basic matching probability to generate the target model includes:

[0096] The source model is adaptively trained using the basic matching probability and the second user feature vector to generate an initial target model; the initial target model has a corresponding target domain.

[0097] The first user feature vector and the second user feature vector are shared to the target domain, and when it is determined that there is a first mapping relationship between the source domain feature vector in the source domain and the target domain feature vector in the target domain, the target domain error value for the target domain is obtained.

[0098] When the target domain error value is less than a preset threshold, the initial target model is determined as the target model.

[0099] In practical applications, generating a target model requires aligning the source and target domains to ensure that the model has a solution. Feature adaptive training is a common method for training models. Its basic idea is to learn common feature representations. In the common feature space, the distributions of the source and target domains should be as similar as possible.

[0100] In this embodiment of the invention, the source model can be trained using the basic matching probability and the second user feature vector to generate an initial target model with a corresponding target domain. Then, the first user feature vector and the second user feature vector are shared to the target domain.

[0101] For example, the features of the recommender in the source model can be shared into the target domain. These features mainly include the recommender's service skill assessment results and basic information about the recommender. The service skill assessment results need to be data from continuous periods, so that new time-series features can be calculated. Alternatively, some user (recommended person) emotional and social relationship features can also be added to the target domain. These features can come from voice or work order information of the user's historical communication with the recommender, as well as the results of the social network analysis model.

[0102] When it is determined that there is a first mapping relationship between the source domain feature vector in the source domain and the target domain feature vector in the target domain, the target domain error value for the target domain is obtained, and when the target domain error value is less than a preset threshold, the initial target model is determined as the target model. That is, in this embodiment of the invention, when the model has a solution, the distance between the edge distribution probabilities of the source domain and the target domain is calculated, and when the distance between the edge distribution probabilities is reached, the target model is determined.

[0103] In practical applications, there may be situations where the model has no solution. For example, due to sparse feature data or strict constraints, the model outputs -1, resulting in an overfitting phenomenon where the user recommendation has no solution.

[0104] Optionally, when it is determined that there is a first mapping relationship between the source domain feature vector in the source domain and the target domain feature vector in the target domain, the training focus can be adjusted by judging whether the source model is overtrained or whether the initial target model is overtrained, and then the source model can be trained until the source domain and the target domain are aligned, so as to ensure that the target model has a solution.

[0105] In this embodiment of the invention, the source model is adaptively trained using the basic matching probability and the second user feature vector to generate an initial target model. The initial target model has a corresponding target domain. The first user feature vector and the second user feature vector are shared to the target domain. When it is determined that there is a first mapping relationship between the source domain feature vector in the source domain and the target domain feature vector in the target domain, the target domain error value for the target domain is obtained. When the target domain error value is less than a preset threshold, the initial target model is determined as the target model. This ensures that the model generates the target model under the premise that there is a solution, further improving the accuracy and efficiency of pushing item information to users through the deep learning model.

[0106] In an optional embodiment of the present invention, it further includes:

[0107] When the target domain error value is greater than or equal to a preset threshold, a second mapping relationship is constructed between the source domain feature vector in the source domain and the target domain feature vector in the target domain; the second mapping relationship is used to adjust the target domain error value.

[0108] In practical applications, if the model's generalization error is below a threshold, the result can be directly output; that is, a target model can be generated to determine the final matching relationship between the recommended party and the recommended item. However, in general, it is not possible to directly obtain the ideal situation where the target domain error is below the threshold. That is, in addition to ensuring the minimization of the source domain error and the alignment of the source and target domains, model denoising is also required.

[0109] In a specific implementation, embodiments of the present invention can construct a second mapping relationship between the source domain feature vector in the source domain and the target domain feature vector in the target domain, and then superimpose it onto the source domain feature vector in the source domain and the target domain feature vector in the target domain. The goal is to reduce the distance between the edge probability distributions of the source domain and the target domain to below a threshold after superimposing the second mapping relationship. The initial target model is then subjected to edge distribution adaptive training until the distance between the edge probability distributions of the source domain and the target domain is reduced to below the threshold.

[0110] In this embodiment of the invention, when the target domain error value is greater than or equal to a preset threshold, a second mapping relationship is constructed between the source domain feature vector in the source domain and the target domain feature vector in the target domain. This ensures that even when the ideal situation of the target domain error being lower than the threshold cannot be directly obtained, the model can be denoised, further improving the accuracy and efficiency of pushing item information to users through the deep learning model.

[0111] To enable those skilled in the art to better understand the embodiments of the present invention, a complete example is provided below to illustrate the embodiments of the present invention.

[0112] refer to Figure 2 , Figure 2 This is a flowchart illustrating an item information push method provided in an embodiment of the present invention.

[0113] Step 1: Develop the source model to effectively constrain the error risk in the source domain, including:

[0114] 1. Extract features and determine the core algorithm of the source model:

[0115] 1) Extracting features from structured data: ① Preprocessing continuous variables, mainly to unify the units of measurement, to ensure that the dimensions do not become invalid due to unit of measurement issues, and to fill in missing values ​​for different dimensions in an appropriate manner; ② Preprocessing discrete variables, mainly time series dimensions, gender, etc., mainly using OneHot encoding.

[0116] 2) Extracting features from unstructured data: ① Building a feature extraction model for text data, which mainly consists of audio recordings, work order descriptions, etc. The Word2Vec method is mainly used for extraction, and the BERT model is mainly used for analysis; ② Building a deep learning feature extraction model for audio data, extracting and labeling features such as tone of voice, emotion, energy, and silence duration.

[0117] 3) Calculate the feature loss rate: ① Input structured data into the source model and calculate the feature loss rate. For those with a loss rate exceeding the threshold, adjust the feature extraction method and fine-tune the model parameters; ② Input unstructured data into the source model and calculate the feature loss rate. For those with a loss rate exceeding the threshold, adjust the feature extraction method and fine-tune the feature extraction model parameters.

[0118] 4) Determine the core algorithm of the source model: ① Select a group of algorithms that can achieve high accuracy and precision after feature learning as the core algorithm of the source model, and set reasonable algorithm selection and integration rules; ② Introduce a multi-task learning model, train the model to understand users based on company revenue and user needs, and output the basic probability of user matching with product.

[0119] 2. Based on the basic probability of user-product matching, extract recommender features using basic recommender information and recommender service skill evaluation results, and increase recommender adaptive training. By adding recommender features and conditions that are easier to obtain than user features, the training cost is reduced. On the one hand, it improves the correlation between tasks, and on the other hand, it transforms user understanding based on historical data into user understanding based on the communication process. During the training process, recommender features need to be reasonably selected to effectively constrain the error risk of the source domain before transfer learning.

[0120] Step 2: Develop a reusable model, align the source and target domains, and ensure the model has solutions.

[0121] Based on the assumption that the scope of recommenders and the multi-task conditions remain unchanged:

[0122] 1. Shared referrer features: The features of the referrer in the source model are shared to the target domain, mainly the service skill assessment results of the referrer and the basic information of the referrer. The service skill assessment results need to be data from continuous stages, so that new time-series features can be calculated in combination.

[0123] 2. Shared user features: Some user emotion and social relationship features are also added to the target domain. The source of these features is mainly voice or ticket information from the user's historical communication with the recommender, as well as the results of the social network analysis model.

[0124] 3. Mapping optimization and solution determination: The mapping of the source domain is optimized through rule superposition and rule transformation. The optimization process focuses on solving the problem of unsolvable recommendation results by filling in the recommender features. That is, the overfitting phenomenon that causes the model to output -1 due to sparse feature data or strict constraints, resulting in no solution for the user recommendation. Since the parameters remain unchanged throughout the training process, further judgment is needed to address the overfitting phenomenon.

[0125] 4. Determine the cause of overfitting and address it: Confirm whether the source model is overtrained or the reused model is overtrained, and then adjust the training focus until the source and target domains are aligned to ensure that the model has a solution.

[0126] Step 3: Model denoising to ultimately determine the matching relationship between users and products:

[0127] 1. Determine if the target domain error is below the threshold: If the model generalization error is below the threshold, the result is directly output, which is the final matching relationship between the user and the product. However, it is generally not possible to directly obtain the ideal situation where the target domain error is below the threshold. That is, in addition to ensuring the minimization of the source domain error and the alignment of the source and target domains, model denoising is also required.

[0128] 2. Superimpose mappings and calculate the distance between marginal probability distributions: To preserve and reuse the training results of the model, a mapping needs to be superimposed on the features of the source and target domains based on the mappings obtained in step two, i.e., ensuring a solution exists. This addresses the problem of excessive generalization error in the model. The calculation aims to reduce the distance between the marginal probability distributions of the source and target domains after mapping to below a threshold.

[0129] 3. Edge Distribution Adaptive Training: Taking the distance between the edge probability distributions of the source domain and the target domain after mapping as the target, edge distribution adaptive training is performed to optimize the superimposed mapping. During this process, the parameters in step two remain unchanged. This method is used to solve for the minimum value of the model's generalization error until the final result is lower than the threshold. At this point, the final one-to-one correspondence between users and products can be obtained, and the product recommendation can be clearly made for the user.

[0130] Through the above methods, 1) by seeking the globally optimal solution for one-to-one matching between products and users, the optimal recommendation strategy output for multi-task recommendation scenarios is achieved, thereby satisfying the interests and demands of multiple parties and realizing multiple improvements in user satisfaction, company development and revenue, and employee satisfaction; 2) through model adaptation, the model achieves good output results in multiple scenarios and multiple tasks, and can output the optimal recommendation strategy for scenarios with different target tasks, applicable to personalized product recommendations across multiple channels such as online and offline; 3) when the number of service personnel is limited, work orders requiring service maintenance can be divided according to the recommendation strategy output by the model, prioritizing user groups with high success rates and high returns, maximizing the efficiency of user segmentation and reaching, and maximizing the business recommendation effect.

[0131] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0132] Reference Figure 3 The diagram shows a structural block diagram of an item information push device provided in an embodiment of the present invention, which may specifically include the following modules:

[0133] The feature vector acquisition module 301 is used to acquire the first user feature vector for the recommended user and the recommended item feature vector for the recommended item;

[0134] The basic matching probability calculation module 302 is used to determine the source model and, based on the source model, calculate and generate the basic matching probability for the recommended user and the recommended item through the first user feature vector and the recommended item feature vector;

[0135] The second user feature vector acquisition module 303 is used to acquire the second user feature vector for the recommender user;

[0136] The target model generation module 304 is used to train the source model using the first user feature vector, the second user feature vector, and the basic matching probability to generate a target model.

[0137] The item information recommendation module 305 is used to push item information for the recommended items using the target model.

[0138] Optionally, the feature vector acquisition module includes:

[0139] The user feature vector acquisition submodule is used to acquire structured and unstructured user feature vectors for the recommended users.

[0140] The item feature vector acquisition submodule is used to obtain structured and unstructured item feature vectors for recommended items.

[0141] Optionally, the basic matching probability calculation module may include:

[0142] The first feature loss rate calculation submodule is used to input the structured user feature vector into the source model and calculate and generate the first feature loss rate;

[0143] The second feature loss rate calculation submodule is used to input the structured recommended item feature vector into the source model and calculate and generate the second feature loss rate.

[0144] The third feature loss rate calculation submodule is used to input the unstructured user feature vector into the source model and calculate the third feature loss rate.

[0145] The fourth feature loss rate calculation submodule is used to input the unstructured recommended item feature vector into the source model and calculate and generate the fourth feature loss rate.

[0146] The basic matching probability calculation submodule is used to calculate and generate a basic matching probability for the recommended user and the recommended item by using the structured user feature vector, the unstructured user feature vector, the structured recommended item feature vector, and the unstructured item feature vector through the source model when the first feature loss rate, the second feature loss rate, the third feature loss rate, and the fourth feature loss rate are lower than a preset threshold.

[0147] Optionally, the source model has a corresponding source domain, and the target model generation module may include:

[0148] An initial target model generation submodule is used to perform feature adaptive training on the source model using the basic matching probability and the second user feature vector to generate an initial target model; the initial target model has a corresponding target domain.

[0149] The target domain error value acquisition submodule is used to share the first user feature vector and the second user feature vector to the target domain, and when it is determined that there is a first mapping relationship between the source domain feature vector in the source domain and the target domain feature vector in the target domain, the target domain error value for the target domain is acquired.

[0150] The target model determination submodule is used to determine the initial target model as the target model when the target domain error value is less than a preset threshold.

[0151] Optionally, it may also include:

[0152] The second mapping relationship construction submodule is used to construct a second mapping relationship between the source domain feature vector in the source domain and the target domain feature vector in the target domain when the target domain error value is greater than or equal to a preset threshold; the second mapping relationship is used to adjust the target domain error value.

[0153] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0154] In addition, this invention also provides an electronic device, including: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the various processes of the above-described item information push method embodiments and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0155] This invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the various processes of the above-described item information push method embodiments and achieves the same technical effects. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0156] Figure 4 A schematic diagram of the hardware structure of an electronic device for implementing various embodiments of the present invention.

[0157] The electronic device 400 includes, but is not limited to, components such as: a radio frequency unit 401, a network module 402, an audio output unit 403, an input unit 404, a sensor 405, a display unit 406, a user input unit 407, an interface unit 408, a memory 409, a processor 410, and a power supply 411. Those skilled in the art will understand that... Figure 4 The electronic device structures shown are not intended to limit the electronic device. An electronic device may include more or fewer components than shown, or combine certain components, or have different component arrangements. In embodiments of the present invention, the electronic device includes, but is not limited to, mobile phones, tablet computers, laptops, PDAs, in-vehicle terminals, wearable devices, and pedometers.

[0158] It should be understood that, in this embodiment of the invention, the radio frequency unit 401 can be used for receiving and transmitting signals during information transmission or calls. Specifically, it receives downlink data from the base station and processes it with the processor 410; additionally, it transmits uplink data to the base station. Typically, the radio frequency unit 401 includes, but is not limited to, an antenna, at least one amplifier, a transceiver, a coupler, a low-noise amplifier, a duplexer, etc. Furthermore, the radio frequency unit 401 can also communicate with networks and other devices through a wireless communication system.

[0159] The electronic device provides users with wireless broadband internet access through network module 402, such as helping users send and receive emails, browse web pages, and access streaming media.

[0160] The audio output unit 403 can convert audio data received by the radio frequency unit 401 or the network module 402 or stored in the memory 409 into audio signals and output them as sound. Furthermore, the audio output unit 403 can also provide audio output related to specific functions performed by the electronic device 400 (e.g., call signal reception sound, message reception sound, etc.). The audio output unit 403 includes a speaker, a buzzer, and a receiver, etc.

[0161] Input unit 404 is used to receive audio or video signals. Input unit 404 may include a graphics processing unit (GPU) 4041 and a microphone 4042. The GPU 4041 processes image data of still images or videos acquired by an image capture device (such as a camera) in video capture mode or image capture mode. The processed image frames can be displayed on display unit 406. The image frames processed by GPU 4041 can be stored in memory 409 (or other storage medium) or transmitted via radio frequency unit 401 or network module 402. Microphone 4042 can receive sound and process such sound into audio data. The processed audio data can be converted into a format that can be transmitted to a mobile communication base station via radio frequency unit 401 in telephone call mode.

[0162] The electronic device 400 also includes at least one sensor 405, such as a light sensor, a motion sensor, and other sensors. Specifically, the light sensor includes an ambient light sensor and a proximity sensor. The ambient light sensor can adjust the brightness of the display panel 4061 according to the ambient light level, and the proximity sensor can turn off the display panel 4061 and / or backlight when the electronic device 400 is moved to the ear. As a type of motion sensor, an accelerometer sensor can detect the magnitude of acceleration in various directions (generally three axes). When stationary, it can detect the magnitude and direction of gravity and can be used to identify the posture of the electronic device (such as landscape / portrait switching, related games, magnetometer posture calibration), vibration recognition related functions (such as pedometer, tapping), etc. The sensor 405 may also include a fingerprint sensor, pressure sensor, iris sensor, molecular sensor, gyroscope, barometer, hygrometer, thermometer, infrared sensor, etc., which will not be described in detail here.

[0163] The display unit 406 is used to display information input by the user or information provided to the user. The display unit 406 may include a display panel 4061, which may be configured in the form of a liquid crystal display (LCD), an organic light-emitting diode (OLED), or the like.

[0164] User input unit 407 can be used to receive input numerical or character information, and generate key signal inputs related to user settings and function control of electronic devices. Specifically, user input unit 407 includes a touch panel 4071 and other input devices 4072. Touch panel 4071, also known as a touch screen, can collect touch operations performed by the user on or near it (such as operations performed by the user using a finger, stylus, or any suitable object or accessory on or near touch panel 4071). Touch panel 4071 may include two parts: a touch detection device and a touch controller. The touch detection device detects the user's touch position and the signal generated by the touch operation, and transmits the signal to the touch controller; the touch controller receives touch information from the touch detection device, converts it into touch point coordinates, and sends it to the processor 410, which receives and executes commands from the processor 410. In addition, touch panel 4071 can be implemented using various types such as resistive, capacitive, infrared, and surface acoustic wave. Besides touch panel 4071, user input unit 407 may also include other input devices 4072. Specifically, other input devices 4072 may include, but are not limited to, physical keyboards, function keys (such as volume control buttons, power buttons, etc.), trackballs, mice, joysticks, etc., which will not be described in detail here.

[0165] Furthermore, the touch panel 4071 can cover the display panel 4061. When the touch panel 4071 detects a touch operation on or near it, it transmits the information to the processor 410 to determine the type of touch event. Subsequently, the processor 410 provides corresponding visual output on the display panel 4061 based on the type of touch event. Although in Figure 4 In this embodiment, the touch panel 4071 and the display panel 4061 are two independent components to realize the input and output functions of the electronic device. However, in some embodiments, the touch panel 4071 and the display panel 4061 can be integrated to realize the input and output functions of the electronic device. The specific implementation is not limited here.

[0166] Interface unit 408 serves as an interface for connecting external devices to electronic device 400. For example, external devices may include a wired or wireless headphone port, an external power supply (or battery charger) port, a wired or wireless data port, a memory card port, a port for connecting a device with an identification module, an audio input / output (I / O) port, a video I / O port, a headphone port, and so on. Interface unit 408 can be used to receive input from external devices (e.g., data, power, etc.) and transmit the received input to one or more components within electronic device 400, or it can be used to transmit data between electronic device 400 and external devices.

[0167] The memory 409 can be used to store software programs and various data. The memory 409 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function (such as sound playback, image playback, etc.), etc.; the data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). Furthermore, the memory 409 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0168] The processor 410 is the control center of the electronic device. It connects various parts of the electronic device via various interfaces and lines. By running or executing software programs and / or modules stored in the memory 409, and by calling data stored in the memory 409, it performs various functions and processes data, thereby providing overall monitoring of the electronic device. The processor 410 may include one or more processing units; preferably, the processor 410 may integrate an application processor and a modem processor. The application processor mainly handles the operating system, user interface, and applications, while the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 410.

[0169] The electronic device 400 may also include a power supply 411 (such as a battery) for supplying power to various components. Preferably, the power supply 411 can be logically connected to the processor 410 through a power management system, thereby enabling functions such as managing charging, discharging, and power consumption through the power management system.

[0170] In addition, the electronic device 400 includes some functional modules not shown, which will not be described in detail here.

[0171] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0172] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0173] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

[0174] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0175] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0176] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0177] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0178] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0179] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0180] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for pushing item information, characterized in that, include: Obtain the first user feature vector for the recommended user and the recommended item feature vector for the recommended item; A source model is determined, and based on the source model, a basic matching probability for the recommended user and the recommended item is calculated using the first user feature vector and the recommended item feature vector; the source model is a multi-task model. Obtain the second user feature vector for the recommender's users; The source model is trained using the first user feature vector, the second user feature vector, and the basic matching probability to generate the target model; The target model is used to push item information for the recommended items; The source model has a corresponding source domain, and the step of training the source model using the first user feature vector, the second user feature vector, and the basic matching probability to generate the target model includes: The source model is subjected to feature adaptive training using the basic matching probability and the second user feature vector to generate an initial target model. The initial target model has a corresponding target domain; The first user feature vector and the second user feature vector are shared to the target domain, and when it is determined that there is a first mapping relationship between the source domain feature vector in the source domain and the target domain feature vector in the target domain, the target domain error value for the target domain is obtained. When the target domain error value is less than a preset threshold, the initial target model is determined as the target model.

2. The method according to claim 1, characterized in that, The steps of obtaining the first user feature vector for the recommended user and the recommended item feature vector for the recommended item include: Obtain structured and unstructured user feature vectors for the recommended users; Obtain the structured feature vector and unstructured feature vector of the recommended items.

3. The method according to claim 2, characterized in that, The step of calculating the basic matching probability for the recommended user and the recommended item based on the source model using the first user feature vector and the recommended item feature vector includes: The structured user feature vector is input into the source model to calculate and generate the first feature loss rate; The structured recommended item feature vector is input into the source model to calculate and generate the second feature loss rate; The unstructured user feature vector is input into the source model to calculate and generate the third feature loss rate; The unstructured recommended item feature vector is input into the source model to calculate and generate the fourth feature loss rate; When the first feature loss rate, the second feature loss rate, the third feature loss rate, and the fourth feature loss rate are all below a preset threshold, the structured user feature vector, the unstructured user feature vector, the structured recommended item feature vector, and the unstructured item feature vector are used to calculate and generate the basic matching probability for the recommended user and the recommended item through the source model.

4. The method according to claim 1, characterized in that, Also includes: When the target domain error value is greater than or equal to a preset threshold, a second mapping relationship is constructed between the source domain feature vector in the source domain and the target domain feature vector in the target domain; the second mapping relationship is used to adjust the target domain error value.

5. An item information push device, characterized in that, include: The feature vector acquisition module is used to acquire the first user feature vector for the recommended user and the recommended item feature vector for the recommended item. The basic matching probability calculation module is used to determine the source model and, based on the source model, calculate the basic matching probability for the recommended user and the recommended item using the first user feature vector and the recommended item feature vector; the source model is a multi-task model. The second user feature vector acquisition module is used to acquire the second user feature vector for the recommender user. The target model generation module is used to train the source model using the first user feature vector, the second user feature vector, and the basic matching probability to generate a target model. The item information recommendation module is used to push item information for the recommended items using the target model; The source model has a corresponding source domain, and the target model generation module includes: An initial target model generation submodule is used to perform feature adaptive training on the source model using the basic matching probability and the second user feature vector to generate an initial target model; the initial target model has a corresponding target domain. The target domain error value acquisition submodule is used to share the first user feature vector and the second user feature vector to the target domain, and when it is determined that there is a first mapping relationship between the source domain feature vector in the source domain and the target domain feature vector in the target domain, the target domain error value for the target domain is acquired. The target model determination submodule is used to determine the initial target model as the target model when the target domain error value is less than a preset threshold.

6. The apparatus according to claim 5, characterized in that, The feature vector acquisition module includes: The user feature vector acquisition submodule is used to acquire structured and unstructured user feature vectors for the recommended users. The item feature vector acquisition submodule is used to obtain structured and unstructured item feature vectors for recommended items.

7. The apparatus according to claim 6, characterized in that, The basic matching probability calculation module includes: The first feature loss rate calculation submodule is used to input the structured user feature vector into the source model and calculate and generate the first feature loss rate; The second feature loss rate calculation submodule is used to input the structured recommended item feature vector into the source model and calculate and generate the second feature loss rate. The third feature loss rate calculation submodule is used to input the unstructured user feature vector into the source model and calculate the third feature loss rate. The fourth feature loss rate calculation submodule is used to input the unstructured recommended item feature vector into the source model and calculate and generate the fourth feature loss rate. The basic matching probability calculation submodule is used to calculate and generate a basic matching probability for the recommended user and the recommended item by using the structured user feature vector, the unstructured user feature vector, the structured recommended item feature vector, and the unstructured item feature vector through the source model when the first feature loss rate, the second feature loss rate, the third feature loss rate, and the fourth feature loss rate are lower than a preset threshold.

8. The apparatus according to claim 5, characterized in that, Also includes: The second mapping relationship construction submodule is used to construct a second mapping relationship between the source domain feature vector in the source domain and the target domain feature vector in the target domain when the target domain error value is greater than or equal to a preset threshold; the second mapping relationship is used to adjust the target domain error value.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes a program stored in the memory, it implements the method as described in any one of claims 1-4.

10. A computer-readable storage medium having instructions stored thereon that, when executed by one or more processors, cause the processors to perform the method as described in any one of claims 1-4.

Images